Convolutional neural network is a good technique for sleep staging based on HRV: A comparative analysis

The fluctuation of heart rate is regulated by autonomic nervous system. In human sleep, the autonomic nervous system plays a leading role. Therefore, we can use heart-rate variability (HRV) to stage the sleep process. Based on two independent public datasets, we construct three end-to-end automatic sleep staging models: fully connected neural networks (FCN), convolutional neural networks (CNN) and long short-term memory networks (LSTM). Only the HRV sequence was used to classify and identify the four sleep stages of the subject's sleep process: wake(W), light sleep (LS), slow-wave sleep (SWS) and rapid eye movement (REM), and the confusion matrix was calculated. The three models were compared by performance index (precision, accuracy, F1, Kappa statistic) and Friedman test. Among these models, the CNN has the best classification effect. The precision of W, REM, LS and SWS were 88.31%, 98.07%, 81.16% and 99.36%, respectively. It's the average accuracy, average F1 value and Kappa statistic were 91.72%, 0.8850 and 0.8844 ± 0.0095, respectively. The experimental results show that the convolutional neural network can achieve good sleep staging effect based on the signal of HRV solely, which is suitable for sleep detection in the home.